Toward fast and accurate human pose estimation via soft-gated skip connections

This paper is on highly accurate and highly efficient human pose estimation.\nRecent works based on Fully Convolutional Networks (FCNs) have demonstrated\nexcellent results for this difficult problem. While residual connections within\nFCNs have proved to be quintessential for achieving high accuracy, we\nre-analyze this design choice in the context of improving both the accuracy and\nthe efficiency over the state-of-the-art. In particular, we make the following\ncontributions: (a) We propose gated skip connections with per-channel learnable\nparameters to control the data flow for each channel within the module within\nthe macro-module. (b) We introduce a hybrid network that combines the HourGlass\nand U-Net architectures which minimizes the number of identity connections\nwithin the network and increases the performance for the same parameter budget.\nOur model achieves state-of-the-art results on the MPII and LSP datasets. In\naddition, with a reduction of 3x in model size and complexity, we show no\ndecrease in performance when compared to the original HourGlass network.\n